Decision, recommendation and causal intelligence.
A score is not permission.
The last engine in the chain decides what to recommend, arbitrates between objectives that disagree, and — where the design supports it — estimates what a change actually caused rather than what merely accompanied it.
Rank, arbitrate, and measure your own effect.
Safe-candidate ranking
Ranking happens only after deterministic filtering has removed everything unsafe. The compiler enforces the order; a recipe that ranks first is rejected.
Objective arbitration
Margin, service level, cash and risk rarely agree. Arbitration makes the trade-off explicit and states which objective won, rather than hiding it in a weighted score.
Uplift
Not who is most likely to convert, but who is most likely to convert because you acted — which is a different and much shorter list.
Experimentation and causal analysis
Exposure recorded, outcomes captured, and a stated identification strategy before any claim about cause is made.
Measuring the effect of your own advice.
A recommendation changes what happens, which contaminates the data you would use to judge the recommendation. The system is built to keep that separable rather than pretending it does not happen.
- Recommendations record intervention exposure — who saw what, and when.
- Optimisation decisions mark their outcomes as influenced, so a result that followed a plan is not counted as evidence the plan was right.
- Process-mined deviations update evidence, never policy.
- Inferred graph links stay separate from asserted edges.
- A mined association cannot be converted into causal wording by a language model.
- Causal conclusions require a defensible identification strategy — a design, not a correlation with a confident sentence attached.
- Ranking occurs only after deterministic safety filtering.
- Action execution always revalidates live state before anything is applied.
A score is not permission.
This is the most important sentence on this page. A high rank means an option scored well against stated objectives — not that it is allowed, and not that it is safe. Permission comes from the deterministic engine and from your live roles, checked again at the moment anything is applied. And a causal claim is held to a stricter standard than a predictive one, because acting on a false cause is more expensive than acting on a weak forecast.
At the end of every recommendation.
- “What should we do about this?” — ranked options, each with the objective it favours and what it costs elsewhere.
- “Which of these customers is worth chasing?” — uplift rather than propensity, so effort goes where it changes the outcome.
- “Did that change work?” — measured against exposure, not against a before-and-after that anything could explain.
- “Why this option over that one?” — the arbitration, stated.
- “What is the next best action here?” — from the safe set only.
Frequently asked questions.
Why is ranking after filtering rather than before?
Because ranking an unsafe option puts it in front of a person, and people act on what they are shown. Filtering first means the unsafe option was never a candidate. The recipe compiler rejects any pipeline that gets this backwards.
What is uplift, and why does it matter?
Propensity tells you who is likely to do something. Uplift tells you who is likely to do it because you acted. Chasing high-propensity customers who would have paid anyway is a well-funded way to achieve nothing.
Can it claim something caused something else?
Only with a defensible identification strategy behind it. A strong correlation and a confident sentence do not qualify, and no language model is allowed to promote an association into a cause.
Does a recommendation ever execute itself?
Only at an autonomy level you have admitted for that specific skill, and even then the action re-resolves permissions and re-checks the fingerprint before applying — refusing the whole request if anything has moved.
See this engine on your own records.
Join the waitlist and ask it something real. Every answer names the engines it used and the records they read.
Join waitlist